GCP – Transforming the contact center with generative AI

Google Cloud has been delivering AI innovations to the contact center for almost a decade. Our Contact Center AI (CCAI) solutions are deployed across almost every industry — from financial services to automotive, retail, healthcare — and especially in telecommunications. Over the course of that time, we have seen tens of millions of call deflections offloading contact center agents and positive effects on call center productivity and customer net promoter scores (NPS).

Telecom was only one of many industries for which our CCAI solutions were intended, but gained immediate traction given the industry’s relatively low industry-wide NPS and desire to improve the customer experience. Now, the breadth of our telecom engagements allows us to start to develop telecom-specific capabilities: prebuilt taxonomies, topic models, virtual agents, human agent assistance, and components and integrations that accelerate deployments. We’ve accumulated expertise across our organizations, including product engineering (to develop unique capabilities aligned to industry use cases), customer engineering to deliver relevant pilots and proofs of concept (PoCs), professional services, and our partner ecosystem, and stand ready to bring that knowledge to the telecom industry.

Specifically, our partners have embraced our CCAI technology for unassisted customer care, addressing a set of frequently asked questions that consume call center agents’ time. Up until recently, topics such as bill explanations, payment arrangements, troubleshooting, and repairs were addressed with deterministic decision trees and probabilistic natural language processing (NLP) in the form of customer support chatbots.

Recently, we took all of our experience and knowledge in delivering AI in the contact center, and extended our tech decisions and methodologies, and began integrating generative AI into our CCAI products and methodologies, at both existing and new customers. Why? Five key reasons. Gen AI:

Allows us to address the broader spectrum of the customer buying journeys in Telecom and beyond, from purchasing decisions to activation to retentionMinimizes the time to value for customers, allowing them to achieve high levels of performance with significantly lower investment — fewer custom models, and deeper integration with unstructured data sources.Improves the development process, shifting from a world of interactive voice response (IVR) and scripted chatbots to a world of intelligent steering and assistive virtual agents.Allows telecoms to pivot from agent offload to agent productivity, providing assisted capabilities that reduce time-to-proficiency and improve agent performanceHelps achieve personalized, proactive, and predictive customer engagement.

Let’s explore how this occurs.

Addressing the customer buying journey

In solutions powered by conversational AI, there have been some use cases that have been off limits, such as queries related to sales or churn (i.e., “retention”). Customers (and technology providers) have generally been conservative about building automated solutions for use cases that drive hard revenue numbers, recognizing that understanding the nuances of customer intent and sentiment, coupled with agent empathy, was too complex to solve deterministically.. In short, virtual agents were insufficiently “human” to understand and react to human emotions.

Large language models (LLMs) have the “superpower” of both explicitly and implicitly understanding intents, including sentiment, and demonstrate emergent reasoning capabilities (especially via Chain of Thought). In fact, Google Cloud’s own research on diagnostic care highlights that LLMs have the ability to be empathetic.

Together, these LLM capabilities have expanded the aperture in which conversational generative agents can engage with customers. As such, user journeys from comparative shopping, to assisted activation, to low-risk retention, to cross-sell / up-sell are all within purview. This inherently shifts the value drivers of CCAI from operational efficiency to revenue realization.

Better integrated with structured and unstructured data

Integrating with complex data sources — structured and unstructured — is key to improving responsiveness and accuracy of LLMs, as well as reducing hallucinations. Two techniques worth noting are retrieval-augmented generation (RAG), and “reason and action” (ReAct) prompting.

RAG effectively retrieves data and augments LLM prompts, giving an LLM access to data that it did not have when it was initially trained. Yet it does not alter the model, ensuring customer privacy and security are maintained.ReAct prompting is a technique that triggers a LLM to reason, or think through (verbally) what it needs to do. A “thoughts,” “actions,” and “observations” approach provides a framework for an LLM to reason through task-specific actions, especially when additional information is retrieved that might be relevant in its reasoning.

Within our Vertex AI Search capabilities, generative agents leverage RAG and other techniques to index your knowledge base across websites, documents, media, image and intranet, and LLMs, both to understand customer queries and intents, as well as to present back data in a conversational way. These agents are designed to be informational in nature. They provide customers with immediate support, while also taking a large amount of top-of-funnel volume off contact center agents and websites.

Gen AI can be used to answer customer or internal use queries on data in documents or websites, but newer multi-modal models can also help create powerful, easy-to-use self-service use cases. For example, customers could take a photo or send an attachment of a problematic bill, and gen AI can understand its key elements (e.g., bill date, bill items) and immediately provide contextual help to the customer (or the human agent trying to help them). For example, imagine a customer trying to diagnose a problem with a broadband router: indexing images and media to allow customers to perform a visual search for similar issues removes friction for the customer — all they need to do is send in a photo of the router.

In a multi-modal gen AI world, customers do not need to understand what is going wrong — they can share what they see of their world and let gen AI interpret what is going wrong based on the CSP’s enterprise data. This helps expand applications of AI beyond customer support to network troubleshooting, field operations, marketing and sales, as well as analytics.

Chatbots to assistive virtual agents

Generative capabilities are directly incorporated into our virtual agent development platform so that we can simultaneously support both highly controlled transactional dialogues, for use cases such as bill payment, while allowing broader sets of diverse questions about a CSP’s business without having to define each and every possible reason for the call. Gen AI is an excellent technology for this, and instead of defining complex dialogues, we can leverage instructions, or playbooks, that provide the necessary steps, in natural language. The same way you would instruct a human agent to follow a runbook, you can instruct gen AI to follow a playbook. This means that use cases that were either too costly or too complex to design can now be implemented with a few lines of natural language instructions.

With these approaches, we can mix and match generative flows and scripted conversational paths across all types of virtual agents — from those that are informational, those that are transactional, and those that are goal-seeking.

From agent offload to agent productivity

While agent offload remains a critical enabler of call center efficiency, the opportunity to improve agent productivity is even larger. Generative techniques can significantly help agent productivity, improving metrics such as average handling time, after call work, ramp-up time, with solutions like summarization and generative knowledge assist that can drive immediate value and are the shortest time to value. Gen AI is foundational to these products — turning transcripts into summaries and knowledge bases into agent recommended responses.

Contact centers are very fragmented, with calls queued to small teams of specialists, and companies invest a lot of money in training and developing those specialists. Generative technology can both take advantage of existing specialists and reduce the need for further specialization by making that expertise available to a much wider audience. Generating recommended responses and providing proactive, real-time guidance is critical to improving overall contact center operational efficiency not just for an agent or a call, but for the overall call center organization. It means less specialization, less sharding of agents, more generalists, and more information on how to resolve issues at the hands of those generalists in real-time. It also means less up-front investment in training, and less risk of sunk costs in the event of high agent churn.

Returning to those agent productivity tools — when we focus on tuning models, tuning prompts, and data preprocessing and postprocessing to specific contact center scenarios — for instance, call transcription and summarization, or topic modeling of call transcripts (as opposed to standard foundation models for summarization or topic identification), we see accuracy, consistency and compliance that leads the industry.

And those transcripts, summaries, and productivity tools can be used to both identify areas of agent development, and for coaching agents, both in the moment and post-call, to improve overall quality.

Personalized, proactive, and predictive customer experiences

Bringing it together, we use product innovations to address to build a more personalized experience for customer interactions, and use AI to help bring more proactivity and prediction to the engagements. There are many real-time and historical signals that allow customer care to be more efficient and more relevant for users, and all of these signals can be absorbed as context for LLMs. Imagine a world where customers interact with your digital touchpoints and your applications use gen AI to provide relevant recommendations based on their customer profile, whether it be in assurance or in sales. When human agents are interacting with their customers, an assistive virtual agent can proactively retrieve customer information from Customer Relationship Management (CRM) or Customer Experience Management (CEM) systems, predict the user’s query in real-time, and have in-the-moment recommendations on how to deal with the customer’s situation, turn by turn, that’s sourced from support documents and websites. Both human and virtual agents can then provide a more personalized experience. Customers are catered to based on their specific needs, leading to more satisfaction and loyalty.

At MWC Barcelona 2024, we’re showcasing demos of how gen AI pulls this all together with our latest product innovations in CCAI Platform, Agent Assist, Insights, and Vertex AI Conversation. Be sure to reach out to your dedicated Google Cloud representative to schedule a demo tour — we look forward to seeing you in Barcelona this year.

For more information on how Google Cloud is partnering with CSPs on their journey to the AI-enabled telco, click here.